Existing unsupervised deep product quantization methods primarily aim for the increased similarity between different views of the identical image, whereas the delicate multi-level semantic similarities preserved between images are overlooked. Moreover, these methods predominantly focus on the Euclidean space for computational convenience, compromising their ability to map the multi-level semantic relationships between images effectively. To mitigate these shortcomings, we propose a novel unsupervised product quantization method dubbed \textbf{Hi}erarchical \textbf{H}yperbolic \textbf{P}roduct \textbf{Q}uantization (HiHPQ), which learns quantized representations by incorporating hierarchical semantic similarity within hyperbolic geometry. Specifically, we propose a hyperbolic product quantizer, where the hyperbolic codebook attention mechanism and the quantized contrastive learning on the hyperbolic product manifold are introduced to expedite quantization. Furthermore, we propose a hierarchical semantics learning module, designed to enhance the distinction between similar and non-matching images for a query by utilizing the extracted hierarchical semantics as an additional training supervision. Experiments on benchmarks show that our proposed method outperforms state-of-the-art baselines.
We present Inferflow, an efficient and highly configurable inference engine for large language models (LLMs). With Inferflow, users can serve most of the common transformer models by simply modifying some lines in corresponding configuration files, without writing a single line of source code. Compared with most existing inference engines, Inferflow has some key features. First, by implementing a modular framework of atomic build-blocks and technologies, Inferflow is compositionally generalizable to new models. Second, 3.5-bit quantization is introduced in Inferflow as a tradeoff between 3-bit and 4-bit quantization. Third, hybrid model partitioning for multi-GPU inference is introduced in Inferflow to better balance inference speed and throughput than the existing partition-by-layer and partition-by-tensor strategies.
Large neural networks require enormous computational clusters of machines. Model-parallel training, when the model architecture is partitioned sequentially between workers, is a popular approach for training modern models. Information compression can be applied to decrease workers communication time, as it is often a bottleneck in such systems. This work explores how simultaneous compression of activations and gradients in model-parallel distributed training setup affects convergence. We analyze compression methods such as quantization and TopK compression, and also experiment with error compensation techniques. Moreover, we employ TopK with AQ-SGD per-batch error feedback approach. We conduct experiments on image classification and language model fine-tuning tasks. Our findings demonstrate that gradients require milder compression rates than activations. We observe that $K=10\%$ is the lowest TopK compression level, which does not harm model convergence severely. Experiments also show that models trained with TopK perform well only when compression is also applied during inference. We find that error feedback techniques do not improve model-parallel training compared to plain compression, but allow model inference without compression with almost no quality drop. Finally, when applied with the AQ-SGD approach, TopK stronger than with $ K=30\%$ worsens model performance significantly.
We present ENTED, a new framework for blind face restoration that aims to restore high-quality and realistic portrait images. Our method involves repairing a single degraded input image using a high-quality reference image. We utilize a texture extraction and distribution framework to transfer high-quality texture features between the degraded input and reference image. However, the StyleGAN-like architecture in our framework requires high-quality latent codes to generate realistic images. The latent code extracted from the degraded input image often contains corrupted features, making it difficult to align the semantic information from the input with the high-quality textures from the reference. To overcome this challenge, we employ two special techniques. The first technique, inspired by vector quantization, replaces corrupted semantic features with high-quality code words. The second technique generates style codes that carry photorealistic texture information from a more informative latent space developed using the high-quality features in the reference image's manifold. Extensive experiments conducted on synthetic and real-world datasets demonstrate that our method produces results with more realistic contextual details and outperforms state-of-the-art methods. A thorough ablation study confirms the effectiveness of each proposed module.
Early wildfire detection in remote and forest areas is crucial for minimizing devastation and preserving ecosystems. Autonomous drones offer agile access to remote, challenging terrains, equipped with advanced imaging technology that delivers both high-temporal and detailed spatial resolution, making them valuable assets in the early detection and monitoring of wildfires. However, the limited computation and battery resources of Unmanned Aerial Vehicles (UAVs) pose significant challenges in implementing robust and efficient image classification models. Current works in this domain often operate offline, emphasizing the need for solutions that can perform inference in real time, given the constraints of UAVs. To address these challenges, this paper aims to develop a real-time image classification and fire segmentation model. It presents a comprehensive investigation into hardware acceleration using the Jetson Nano P3450 and the implications of TensorRT, NVIDIA's high-performance deep-learning inference library, on fire classification accuracy and speed. The study includes implementations of Quantization Aware Training (QAT), Automatic Mixed Precision (AMP), and post-training mechanisms, comparing them against the latest baselines for fire segmentation and classification. All experiments utilize the FLAME dataset - an image dataset collected by low-altitude drones during a prescribed forest fire. This work contributes to the ongoing efforts to enable real-time, on-board wildfire detection capabilities for UAVs, addressing speed and the computational and energy constraints of these crucial monitoring systems. The results show a 13% increase in classification speed compared to similar models without hardware optimization. Comparatively, loss and accuracy are within 1.225% of the original values.
The emergence of accurate open large language models (LLMs) has led to a race towards performant quantization techniques which can enable their execution on end-user devices. In this paper, we revisit the problem of "extreme" LLM compression-defined as targeting extremely low bit counts, such as 2 to 3 bits per parameter-from the point of view of classic methods in Multi-Codebook Quantization (MCQ). Our algorithm, called AQLM, generalizes the classic Additive Quantization (AQ) approach for information retrieval to advance the state-of-the-art in LLM compression, via two innovations: 1) learned additive quantization of weight matrices in input-adaptive fashion, and 2) joint optimization of codebook parameters across each transformer blocks. Broadly, AQLM is the first scheme that is Pareto optimal in terms of accuracy-vs-model-size when compressing to less than 3 bits per parameter, and significantly improves upon all known schemes in the extreme compression (2bit) regime. In addition, AQLM is practical: we provide fast GPU and CPU implementations of AQLM for token generation, which enable us to match or outperform optimized FP16 implementations for speed, while executing in a much smaller memory footprint.
Quantization (Alistarh et al., 2017) is an important (stochastic) compression technique that reduces the volume of transmitted bits during each communication round in distributed model training. Suresh et al. (2022) introduce correlated quantizers and show their advantages over independent counterparts by analyzing distributed SGD communication complexity. We analyze the forefront distributed non-convex optimization algorithm MARINA (Gorbunov et al., 2022) utilizing the proposed correlated quantizers and show that it outperforms the original MARINA and distributed SGD of Suresh et al. (2022) with regard to the communication complexity. We significantly refine the original analysis of MARINA without any additional assumptions using the weighted Hessian variance (Tyurin et al., 2022), and then we expand the theoretical framework of MARINA to accommodate a substantially broader range of potentially correlated and biased compressors, thus dilating the applicability of the method beyond the conventional independent unbiased compressor setup. Extensive experimental results corroborate our theoretical findings.
Diffusion models have achieved great success in image generation tasks through iterative noise estimation. However, the heavy denoising process and complex neural networks hinder their low-latency applications in real-world scenarios. Quantization can effectively reduce model complexity, and post-training quantization (PTQ), which does not require fine-tuning, is highly promising for compressing and accelerating diffusion models. Unfortunately, we find that due to the highly dynamic distribution of activations in different denoising steps, existing PTQ methods for diffusion models suffer from distribution mismatch issues at both calibration sample level and reconstruction output level, which makes the performance far from satisfactory, especially in low-bit cases. In this paper, we propose Enhanced Distribution Alignment for Post-Training Quantization of Diffusion Models (EDA-DM) to address the above issues. Specifically, at the calibration sample level, we select calibration samples based on the density and variety in the latent space, thus facilitating the alignment of their distribution with the overall samples; and at the reconstruction output level, we modify the loss of block reconstruction with the losses of layers, aligning the outputs of quantized model and full-precision model at different network granularity. Extensive experiments demonstrate that EDA-DM significantly outperforms the existing PTQ methods across various models (DDIM, LDM-4, LDM-8, Stable-Diffusion) and different datasets (CIFAR-10, LSUN-Bedroom, LSUN-Church, ImageNet, MS-COCO).
Face image synthesis is gaining more attention in computer security due to concerns about its potential negative impacts, including those related to fake biometrics. Hence, building models that can detect the synthesized face images is an important challenge to tackle. In this paper, we propose a fusion-based strategy to detect face image synthesis while providing resiliency to several attacks. The proposed strategy uses a late fusion of the outputs computed by several undisclosed models by relying on random polynomial coefficients and exponents to conceal a new feature space. Unlike existing concealing solutions, our strategy requires no quantization, which helps to preserve the feature space. Our experiments reveal that our strategy achieves state-of-the-art performance while providing protection against poisoning, perturbation, backdoor, and reverse model attacks.
We investigate parameter-efficient fine-tuning (PEFT) methods that can provide good accuracy under limited computational and memory budgets in the context of large language models (LLMs). We present a new PEFT method called Robust Adaptation (RoSA) inspired by robust principal component analysis that jointly trains $\textit{low-rank}$ and $\textit{highly-sparse}$ components on top of a set of fixed pretrained weights to efficiently approximate the performance of a full-fine-tuning (FFT) solution. Across a series of challenging generative tasks such as grade-school math and SQL query generation, which require fine-tuning for good performance, we show that RoSA outperforms LoRA, pure sparse fine-tuning, and alternative hybrid methods at the same parameter budget, and can even recover the performance of FFT on some tasks. We provide system support for RoSA to complement the training algorithm, specifically in the form of sparse GPU kernels which enable memory- and computationally-efficient training, and show that it is also compatible with low-precision base weights, resulting in the first joint representation combining quantization, low-rank and sparse approximations. Our code is available at https://github.com/IST-DASLab/RoSA.
The emergence of billion-parameter diffusion models such as Stable Diffusion XL, Imagen, and DALL-E 3 has significantly propelled the domain of generative AI. However, their large-scale architecture presents challenges in fine-tuning and deployment due to high resource demands and slow inference speed. This paper explores the relatively unexplored yet promising realm of fine-tuning quantized diffusion models. Our analysis revealed that the baseline neglects the distinct patterns in model weights and the different roles throughout time steps when finetuning the diffusion model. To address these limitations, we introduce a novel memory-efficient fine-tuning method specifically designed for quantized diffusion models, dubbed TuneQDM. Our approach introduces quantization scales as separable functions to consider inter-channel weight patterns. Then, it optimizes these scales in a timestep-specific manner for effective reflection of the role of each time step. TuneQDM achieves performance on par with its full-precision counterpart while simultaneously offering significant memory efficiency. Experimental results demonstrate that our method consistently outperforms the baseline in both single-/multi-subject generations, exhibiting high subject fidelity and prompt fidelity comparable to the full precision model.
This paper presents a video coding scheme that combines traditional optimization methods with deep learning methods based on the Enhanced Compression Model (ECM). In this paper, the traditional optimization methods adaptively adjust the quantization parameter (QP). The key frame QP offset is set according to the video content characteristics, and the coding tree unit (CTU) level QP of all frames is also adjusted according to the spatial-temporal perception information. Block importance mapping technology (BIM) is also introduced, which adjusts the QP according to the block importance. Meanwhile, the deep learning methods propose a convolutional neural network-based loop filter (CNNLF), which is turned on/off based on the rate-distortion optimization at the CTU and frame level. Besides, intra-prediction using neural networks (NN-intra) is proposed to further improve compression quality, where 8 neural networks are used for predicting blocks of different sizes. The experimental results show that compared with ECM-3.0, the proposed traditional methods and adding deep learning methods improve the PSNR by 0.54 dB and 1 dB at 0.05Mbps, respectively; 0.38 dB and 0.71dB at 0.5 Mbps, respectively, which proves the superiority of our method.
This paper revisits the event-triggered control problem from a data-driven perspective, where unknown continuous-time linear systems subject to disturbances are taken into account. Using data information collected off-line instead of accurate system model information, a data-driven dynamic event-triggered control scheme is developed in this paper. The dynamic property is reflected by that the designed event-triggering function embedded in the event-triggering mechanism (ETM) is dynamically updated as a whole. Thanks to this dynamic design, a strictly positive minimum inter-event time (MIET) is guaranteed without sacrificing control performance. Specifically, exponential input-to-state stability (ISS) of the closed-loop system with respect to disturbances is achieved in this paper, which is superior to some existing results that only guarantee a practical exponential ISS property. The dynamic ETM is easy-to-implement in practical operation since all designed parameters are determined only by a simple data-driven linear matrix inequality (LMI), without additional complicated conditions as required in relevant literature. As quantization is the most common signal constraint in practice, the developed control scheme is further extended to the case where state transmission is affected by a uniform or logarithmic quantization effect. Finally, adequate simulations are performed to show the validity and superiority of the proposed control schemes.
Transformer-based Large Language Models (LLMs) have made a significant impact on various domains. However, LLMs' efficiency suffers from both heavy computation and memory overheads. Compression techniques like sparsification and quantization are commonly used to mitigate the gap between LLM's computation/memory overheads and hardware capacity. However, existing GPU and transformer-based accelerators cannot efficiently process compressed LLMs, due to the following unresolved challenges: low computational efficiency, underutilized memory bandwidth, and large compilation overheads. This paper proposes FlightLLM, enabling efficient LLMs inference with a complete mapping flow on FPGAs. In FlightLLM, we highlight an innovative solution that the computation and memory overhead of LLMs can be solved by utilizing FPGA-specific resources (e.g., DSP48 and heterogeneous memory hierarchy). We propose a configurable sparse DSP chain to support different sparsity patterns with high computation efficiency. Second, we propose an always-on-chip decode scheme to boost memory bandwidth with mixed-precision support. Finally, to make FlightLLM available for real-world LLMs, we propose a length adaptive compilation method to reduce the compilation overhead. Implemented on the Xilinx Alveo U280 FPGA, FlightLLM achieves 6.0$\times$ higher energy efficiency and 1.8$\times$ better cost efficiency against commercial GPUs (e.g., NVIDIA V100S) on modern LLMs (e.g., LLaMA2-7B) using vLLM and SmoothQuant under the batch size of one. FlightLLM beats NVIDIA A100 GPU with 1.2$\times$ higher throughput using the latest Versal VHK158 FPGA.
Recently, high-fidelity scene reconstruction with an optimized 3D Gaussian splat representation has been introduced for novel view synthesis from sparse image sets. Making such representations suitable for applications like network streaming and rendering on low-power devices requires significantly reduced memory consumption as well as improved rendering efficiency. We propose a compressed 3D Gaussian splat representation that utilizes sensitivity-aware vector clustering with quantization-aware training to compress directional colors and Gaussian parameters. The learned codebooks have low bitrates and achieve a compression rate of up to $31\times$ on real-world scenes with only minimal degradation of visual quality. We demonstrate that the compressed splat representation can be efficiently rendered with hardware rasterization on lightweight GPUs at up to $4\times$ higher framerates than reported via an optimized GPU compute pipeline. Extensive experiments across multiple datasets demonstrate the robustness and rendering speed of the proposed approach.
Transformer has been adopted to image recognition tasks and shown to outperform CNNs and RNNs while it suffers from high training cost and computational complexity. To address these issues, a hybrid approach has become a recent research trend, which replaces a part of ResNet with an MHSA (Multi-Head Self-Attention). In this paper, we propose a lightweight hybrid model which uses Neural ODE (Ordinary Differential Equation) as a backbone instead of ResNet so that we can increase the number of iterations of building blocks while reusing the same parameters, mitigating the increase in parameter size per iteration. The proposed model is deployed on a modest-sized FPGA device for edge computing. The model is further quantized by QAT (Quantization Aware Training) scheme to reduce FPGA resource utilization while suppressing the accuracy loss. The quantized model achieves 79.68% top-1 accuracy for STL10 dataset that contains 96$\times$96 pixel images. The weights of the feature extraction network are stored on-chip to minimize the memory transfer overhead, allowing faster inference. By eliminating the overhead of memory transfers, inference can be executed seamlessly, leading to accelerated inference. The proposed FPGA implementation accelerates the backbone and MHSA parts by 34.01$\times$, and achieves an overall 9.85$\times$ speedup when taking into account the software pre- and post-processing. The FPGA acceleration leads to 7.10$\times$ better energy efficiency compared to the ARM Cortex-A53 CPU. The proposed lightweight Transformer model is demonstrated on Xilinx ZCU104 board for the image recognition of 96$\times$96 pixel images in this paper and can be applied to different image sizes by modifying the pre-processing layer.
The blooming of social media and face recognition (FR) systems has increased people's concern about privacy and security. A new type of adversarial privacy cloak (class-universal) can be applied to all the images of regular users, to prevent malicious FR systems from acquiring their identity information. In this work, we discover the optimization dilemma in the existing methods -- the local optima problem in large-batch optimization and the gradient information elimination problem in small-batch optimization. To solve these problems, we propose Gradient Accumulation (GA) to aggregate multiple small-batch gradients into a one-step iterative gradient to enhance the gradient stability and reduce the usage of quantization operations. Experiments show that our proposed method achieves high performance on the Privacy-Commons dataset against black-box face recognition models.
Quantization is of significance for compressing the over-parameterized deep neural models and deploying them on resource-limited devices. Fixed-precision quantization suffers from performance drop due to the limited numerical representation ability. Conversely, mixed-precision quantization (MPQ) is advocated to compress the model effectively by allocating heterogeneous bit-width for layers. MPQ is typically organized into a searching-retraining two-stage process. In this paper, we devise a one-shot training-searching paradigm for mixed-precision model compression. Specifically, in the first stage, all potential bit-width configurations are coupled and thus optimized simultaneously within a set of shared weights. However, our observations reveal a previously unseen and severe bit-width interference phenomenon among highly coupled weights during optimization, leading to considerable performance degradation under a high compression ratio. To tackle this problem, we first design a bit-width scheduler to dynamically freeze the most turbulent bit-width of layers during training, to ensure the rest bit-widths converged properly. Then, taking inspiration from information theory, we present an information distortion mitigation technique to align the behavior of the bad-performing bit-widths to the well-performing ones. In the second stage, an inference-only greedy search scheme is devised to evaluate the goodness of configurations without introducing any additional training costs. Extensive experiments on three representative models and three datasets demonstrate the effectiveness of the proposed method. Code can be available on \href{https://www.github.com/1hunters/retraining-free-quantization}{https://github.com/1hunters/retraining-free-quantization}.
We present a framework for generating full-bodied photorealistic avatars that gesture according to the conversational dynamics of a dyadic interaction. Given speech audio, we output multiple possibilities of gestural motion for an individual, including face, body, and hands. The key behind our method is in combining the benefits of sample diversity from vector quantization with the high-frequency details obtained through diffusion to generate more dynamic, expressive motion. We visualize the generated motion using highly photorealistic avatars that can express crucial nuances in gestures (e.g. sneers and smirks). To facilitate this line of research, we introduce a first-of-its-kind multi-view conversational dataset that allows for photorealistic reconstruction. Experiments show our model generates appropriate and diverse gestures, outperforming both diffusion- and VQ-only methods. Furthermore, our perceptual evaluation highlights the importance of photorealism (vs. meshes) in accurately assessing subtle motion details in conversational gestures. Code and dataset available online.
Consider a linear quadratic regulator (LQR) problem being solved in a model-free manner using the policy gradient approach. If the gradient of the quadratic cost is being transmitted across a rate-limited channel, both the convergence and the rate of convergence of the resulting controller may be affected by the bit-rate permitted by the channel. We first pose this problem in a communication-constrained optimization framework and propose a new adaptive quantization algorithm titled Adaptively Quantized Gradient Descent (AQGD). This algorithm guarantees exponentially fast convergence to the globally optimal policy, with no deterioration of the exponent relative to the unquantized setting, above a certain finite threshold bit-rate allowed by the communication channel. We then propose a variant of AQGD that provides similar performance guarantees when applied to solve the model-free LQR problem. Our approach reveals the benefits of adaptive quantization in preserving fast linear convergence rates, and, as such, may be of independent interest to the literature on compressed optimization. Our work also marks a first step towards a more general bridge between the fields of model-free control design and networked control systems.